HF Model & Benchmark Matrix

24,078 variants · 14,984 scores · 120 benchmarks

Qwen3.5-35B-A3B

Qwen/Qwen3.5-35B-A3B

kind
original
quant
bf16 (16-bit)
family
qwen
modality
image-text-to-text
root model
Qwen/Qwen3.5-35B-A3B-Base
provider
Qwen
license
apache-2.0
language
—
context
262,144
library
transformers
vRAM (GB)
—
size on disk
—
downloads
2,537,450
likes
1,476
created
—
last modified
2026-04-24T02:38:38+00:00 (2026)
huggingface
https://huggingface.co/Qwen/Qwen3.5-35B-A3B
tags
transformerssafetensorsqwen3_5_moeimage-text-to-textconversationalbase_model:Qwen/Qwen3.5-35B-A3B-Basebase_model:finetune:Qwen/Qwen3.5-35B-A3B-Baselicense:apache-2.0eval-resultsendpoints_compatibledeploy:azureregion:us

Benchmark scores (0)

no scores extracted for this variant.

Discussions (82)

titleauthorstatuscommentstype
Question about M-RoPE section: why not allocate more rotary budget to the temporal axis?vBaiCaiopen0disc
merge -> Agents-A1 / Ornith-1.0-35B / Qwen3.5-35B-A3Btepiraleopen0disc
Listed on OpenModelMapduola15open0disc
Listed on OpenModelMapduola15open0disc
Listed on OpenModelMapduola15open0disc
Add EvalEval community eval resultsEvalEvalBotopen0disc
Given a search question, return reasoning and content("\n\n") with no tool calls in first turn ending with finish_reason==stop.xingmoopen0disc
Add MMMU-Pro evaluation resultSaylorTwiftopen0disc
When will there be an Instruct version?aklghopen0disc
[Docs] Conflicting sampling parameters in README.mdtrgtszropen0disc
Add vLLM offline inference example for Qwen3.5 -- Added in README.mdhrithiksagar-bgenopen0disc
Compact tool-rendering template — 49% fewer prompt tokensEvitaopen0disc
[Bug] Jinja template preventing multiple system messagesphly95open0disc
Reddit Claim - training bug in Qwen3.5 35B A3Bjpsequeiraopen0disc
Some tokens in Qwen3.5 vocabulary cannot be generated from the merges filejslhclopen0disc
fix chat template to avoid empty historical `<think>` blockslatent-variableopen0disc
Why does the thinking mode in Qwen3.5 35B MoE only output a closing </think> tag without an opening <think> tag?Ttooongopen0disc
Qwen3.5 - 35b some time skip the thinking stage then I don't disable thinkingming514810open0disc
Non-deterministic outputs with identical inputs (GPTQ-Int4, vLLM 0.18, document extraction)Starlight1997open0disc
Question about tool definition format in chat template: should tools be passed as JSON or XML?SanJiaoMaoopen0disc
First MTP GGUF + custom quantization pipeline for this model — looking for arXiv endorserKevletesteuropen0disc
llama.cpp load issue with this model.. GGUFLinuxMagicopen0disc
Fixed jinja template that was causing stalling and stopping the responses when using tools.deladuckopen0disc
Proposal: new chat_template_arg `enable_history_reasoning` for reusing prompt cache among querys within Agents .Abioyopen0disc
Add ScreenSpot-Pro evaluation resultmerveopen0disc
Add MathArena evaluation result for hmmt/hmmt_feb_2026JasperDekoninckopen0disc
Add MathArena evaluation result for aime/aime_2026JasperDekoninckopen0disc
21-fix chat template — addresses tool calling crash (discussion #4) and other open bugsdeletedopen0disc
Hugging Face vision model using questionDay1Kimopen0disc
源神!DGHMFFDopen0disc
Questiony198open0disc
I paid $3,800 to build me a 30-agent AI swarm. It then moved into my house and stopped costing me money.digiegoldopen0disc
Update README.mdcryptosxionopen0disc
uhcryptosxionopen0disc
llama.cpp prompt reanalyze issuemayankiit04open0disc
Forget this nonsnseZiggySopen0disc
hereDogger77closed0disc
SafetyProfG5open0disc
goodpallabkr77open0disc
RIPkabachuhaopen0disc
Benchmarks, tradeoffs, and live video inference across five Qwen 3.5 modelsyounes-ovsopen0disc
Learning to be ableFairyydr0popen0disc
Model feedback: Problem with handling latest new articlesbabytifaopen0disc
Upload ChromeSetup.execabot1open0disc
vLLM - Looping preventionjanreges3open0disc
One flaw in the architecture - has to reprocess the prompt every timeDampfinchenopen0disc
hello everyonehoangminhkhoiopen0disc
tempekFunkistopen0disc
wwwwiiamyeraopen0disc
dgddhhdeafolarbi1closed0disc
TestingMoodymueopen0disc
Upload test.txtMoodymueopen0disc
TitleReimotionopen0disc
can not get correct outputchelvanclosed0disc
Qwen Team :- Qwen3-coder-next vs Qwen3.5-35B-A3B for codingmayankiit04open0disc
comment here if you are from actionfi hugging face taskjairuslopen0disc
ghuonjayengkoploopen0disc
CHAMPIONS LEAGUEdreezy05open0disc
wow coolprincejoe12open0disc
kallyhVarij179open0disc
Is `qwen3_nonthinking.jinja` available for disabling thinking?kraftDongopen0disc
just a bit of adjustment in the textamoduadeolaclosed0disc
Update README.mdamoduadeolaopen0disc
My suggestionamoduadeolaopen0disc
Reasoning content leaks into `message.content` with JSON schema response formatTikhonumopen0disc
Quantization AWQ INT4abbas381366open0disc
I would like a recommended training environment setup for the Qwen3.5-MoE model (e.g., Qwen3.5-35B-A3B, model_type: qwen3_5_moe).444515liuxinopen0disc
数字和中文之间多了空格,在某些场景下完全用不了。waxwax0099open0disc
95tgfcajayyxzclosed0disc
jsjsriofbrn666closed0disc
Update README.mdSadiqeenopen0disc
Add evaluation resultsSaylorTwiftopen0disc
Context-Folding Reproducibility | 可复现性 ?pandemoopen0disc
Chat template for Pi AGentnimishchaudhariopen0disc
LESS GOoAkicouopen0disc
🚀 Qwen3.5 Dense/MoE Training Supportstudy-hjtopen0disc
Installation Video and Testing - Step by Stepfahdmirzacopen0disc
Why is this part of the default template?DaGhostmanopen0disc
tool calling chat template is brokenfullstackopen0disc
Instruct mode metricsBaskermasopen0disc
Great release! Thank you Qwen!Dampfinchenopen0disc
请千万千万别忘了把 Qwen Image 2.0 也开源——这对我们本地用户来说会是个巨大的改变 :-)Hanswalteropen0disc
Hugging Face model & benchmark data scraped & matrix-ified. 24,078 variants · 14,984 scores · 120 benchmarks.